Drum Transcription Using Automatic Grouping of Events and Prior Subspace Analysis

While Prior Subspace Analysis (PSA) has proved an effective tool for transcribing mixtures of snare, kick drum and hi-hat both in the “drums-only” case and in the presence of pitched instruments attempts to extend it to deal with increased numbers of drum types have met with mixed results. To overcome this an automatic modeling and grouping procedure has been developed which groups drum events on the similarity of their frequency content. Combining this procedure with PSA allows the extension of PSA to robustly handle greater numbers of drum types. The effectiveness of this approach is demonstrated in a drum transcription algorithm. 1. PRIOR SUBSPACE ANALYSIS Prior Subspace Analysis (PSA) is a technique for sound source separation in single channel mixtures in cases where prior knowledge is available about the sources [1,2]. PSA represents sound sources as low dimensional independent subspaces in the timefrequency plane and is based on Independent Subspace Analysis (ISA) and the generalised sound classification techniques created by Casey [3,4]. It uses prior knowledge about the sources to overcome a number of problems associated with ISA not least of which is the problem of estimating the amount of information to be retained from the dimensional reduction stage of ISA. The mixture signal is transferred to a timefrequency representation such as a spectrogram. PSA then assumes that the overall spectrogram Y results from the summation of l unknown independent spectrograms Yj. This yields